JP2004191968A - Method and device for separating signal source - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a technique for separating a signal from a mixture of a 1st source signal related to a 1st source and a 2nd source signal related to a 2nd source. <P>SOLUTION: Two signals are obtained first which represent two mixtures of the 1st source signal and 2nd source signal. Those two signals and at least one known statistical characteristic related to the 1st and 2nd sources are used and the 1st source signal is separated from the mixtures in a nonlinear signal domain without the need to use a reference signal. <P>COPYRIGHT: (C)2004,JPO&NCIPI

Description

The present invention relates generally to signal separation techniques, and more particularly, if some statistical property is known for each source, e.g., the probability density function of each source is a known Gaussian mixture. of the Gaussians, which relates to techniques for separating non-linear mixtures of sources.

  Source separation addresses the problem of recovering these source signals by observing different mixtures of the source signals. The usual approach to source separation generally assumes that the source signals are mixed linearly. Also, the usual method for source separation is that no detailed information about the statistical properties of the source is known (or little detail in the semi-blind method) and that explicit Is generally blind in the sense that it is assumed that it can be used for The method disclosed in JF Cardoso's paper entitled "Blind SignalSeparation: Statistical Principles" in Proceedings of the IEEE, vol. 9, October 1998, pp. 2009-2025, assumes a linear mixture and is blind. Is an example of the source separation method.

The method disclosed in the paper by A. Acero et al., Entitled "Speech / Noise Separation Using Two Microphones and a VQ Model of Speech Signals," Proceedings of ICSLP 2000, uses an a priori approach to the source probability density function (pdf). We propose a source separation technique that uses sensitive information. However, since the technique operates in the Linear Predictive Coefficient (LPC) domain due to the linear transformation of the waveform domain, the technique assumes that the observed mixture is linear. Therefore, the technique cannot be used in the case of non-linear mixing.

  However, in some cases, the observed mixture is not linear, and in some cases, a priori information regarding the statistical properties of the source can be obtained with high reliability. This is the case, for example, in audio applications that require the separation of mixed audio sources. Examples of such speech applications are speech recognition in the presence of competing speech, interfering music, or special noise sources, such as car or street noise.

  Even though the audio source may be assumed to be linearly mixed in the waveform domain, linear mixing of the waveforms results in non-linear mixing in the cepstral domain, the domain where speech applications typically operate. As is known, cepstra is a vector calculated by the front end of a speech recognition system from the log spectrum of a segment of a speech waveform. See, for example, an article by L. Rabiner et al. In the PrenticeHall Signal Processing Series, 1993, entitled "Fundamentals of Speech Recognition" chapter 3.

  Because of this log transformation, the linear mixing of the waveform signals results in the non-linear mixing of the cepstral signals. However, it is computationally advantageous in speech applications to perform source separation in the cepstral domain than in the waveform domain. In fact, a Sepstra stream corresponding to the generated sound is calculated from continuously superimposed segments of the audio waveform. A segment is typically about 100 milliseconds (ms) long, and the shift between two adjacent segments is about 10 ms long. Thus, a separation process that operates on 11 kilohertz (kHz) audio data in the cepstral domain needs to be applied every 110 samples, as compared to the waveform domain where the separation process must be applied to each sample. Only.

  Further, the pdf of speech and the pdf of many possible interfering audio signals (eg, competing speech, music, particular noise sources, etc.) can be reliably modeled in the cepstral domain and integrated in the separation process. . The pdf of the speech in the cepstral domain is calculated for recognition purposes, and the pdf of the interference source may be calculated off-line for a representative set of data collected from similar sources.

  The method disclosed by S. Deligne and R. Gopinath, entitled "RobustSpeech Recognition with Multi-channel Codebook Dependent Cepstral Normalization (MCDCN)" in Proceedings of ASRU 2001 and 2002, is a method for pdfs of at least one source. We propose a source separation technique that integrates experimental information and does not assume linear mixing. In this method, unwanted source signals interfere with desired source signals. A mixture of the desired signal and the interfering signal is recorded on one channel, while only the interfering signal (ie, not containing the desired signal) is recorded on a second channel forming a so-called reference signal. However, in many cases, the reference signal is not available. For example, in the context of a car speech recognition application and the competing speech of a car passenger, it is not possible to separately capture the speech of the user (eg, driver) of the speech recognition system and the competing speech of other passengers in the car. It is possible.

Accordingly, there is a need for a source separation technique that overcomes the disadvantages and disadvantages associated with conventional source separation techniques.
A paper by JF Cardoso entitled "Blind Signal Separation: StatisticalPrinciples" in Proceedings of the IEEE, vol. 9, October 1998, pp. 2009-2025. Proceedings of ICSLP2000, a paper by A. Acero and others entitled "Speech / Noise Separation Using Two Microphones and a VQ Model of SpeechSignals." A paper by L. Rabiner et al. Entitled Chapter 3, "Fundamentals of Speech Recognition", Prentice Hall SignalProcessing Series, 1993. A paper by S. Deligne and R. Gopinath entitled "Robust Speech Recognition with Multi-channel Codebook Dependent Cepstral Normalization (MCDCN)" in Proceedings of ASRU 2001 and 2002.

  It is an object of the present invention to provide an improved speech separation technique.

  In one aspect of the invention, a technique for separating a signal from a mixture of a first source signal associated with a first source and a second source signal associated with a second source comprises the following steps / operations. Including. First, two mixed signals respectively representing two mixtures of the first source signal and the second source signal are obtained. Thus, using the two mixed signals and at least one known statistical property associated with the first and second sources, and without requiring the use of a reference signal, in the nonlinear signal domain, A first source signal is separated from the mixture.

  The two resulting mixed signals represent an unweighted mixed signal of the first and second source signals and a weighted mixed signal of the first and second source signals, respectively. The separation step / operation may be performed in the non-linear domain by converting the unweighted mixed signal to a first cepstral mixed signal and converting the weighted mixed signal to a second cepstral mixed signal.

  Accordingly, the separating step / operation further comprises iteratively generating an estimate for the second source signal based on the estimate for the second cepstral mixed signal and the first source signal from a previous iteration in the separating step / operation. May be included. Preferably, the step / operation of generating an estimate for the second source signal assumes that the second source signal is modeled by Gaussian mixing.

  Further, the separating step / operation may include iteratively generating an estimate for the first source signal based on the estimate for the first cepstral mixed signal and the second source signal. Preferably, the step / operation of generating an estimate for the first source signal assumes that the first source signal is modeled by Gaussian mixing.

  After the separation process, the separated first source signal may be subsequently used by a signal processing application, for example, a speech recognition application. Further, in some audio processing applications, the first source signal may be an audio signal and the second source signal may be a signal representing competing audio, interfering music, and a particular noise source. .

  These and other objects, features, and advantages of the present invention will become apparent from the following detailed description of illustrative embodiments of the invention, which should be read in conjunction with the accompanying drawings.

  The present invention is described below in connection with an exemplary speech recognition application. Further, the exemplary speech recognition application is considered "codebook dependent." It should be understood that the phrase "codebook-dependent" refers to using Gaussian mixture to model the probability density function of each source signal. The codebook associated with a source signal includes a set of codewords that characterize the source signal. Each codeword is specified by its prior probabilities and by the parameters of the Gaussian distribution, namely the mean matrix and the covariance matrix. In other words, Gaussian mixture is the same as codebook.

  However, it should be further understood that the invention is not limited to this application and any particular application. Rather, the present invention does not assume linear mixing of the sources, assumes that at least one statistical characteristic of the sources is known, and is desirable to perform a source separation process that does not require a reference signal. More generally applicable to applications.

  Therefore, before describing the source separation process of the present invention in the context of speech recognition, first the principles of the source separation of the present invention will be generally described.

  ypcm1 and ypcm2 are two waveform signals that are linearly mixed, so that the two mixed xpcm1 and xpcm2 result according to xpcm1 = ypcm1 + ypcm2 and xpcm2 = a ypcm1 + ypcm2, where a <1. Assume. Further assume that yf1 and yf2 are the spectra of signals ypcm1 and ypcm2, respectively, and xf1 and xf2 are the spectra of signals xpcm1 and xpcm2, respectively.

Further, y1, y2, x1 and x2 are respectively y1 = Clog (yf1), y2 = C log (yf2), x1 = C log (xf1), and x2 = C log (xf2) according to yf1, yf2, xf1, This is a cepstral signal corresponding to xf2. Note that C indicates a discrete cosine transform. Therefore, the following equation is shown:
y1 = x1-g (y1, y2,1) (1)
y2 = x2-g (y2, y1, a) (2)
Note that g (u, v, w) = C log (1 + wexp (invC (vu))), and invC indicates an inverse discrete cosine transform.

Since y1 in equation (1) is unknown, the value of the function is approximated by its predicted value beyond y1, ie, Ey1 [g (y1, y2,1) | y2]. Where the predicted value is calculated for a Gaussian mixture that models the pdf of y1. Also, since y2 in equation (2) is also unknown, the value of function g is approximated by its predicted value exceeding y2, ie, Ey2 [g (y2, y1, a) | y1]. Where the predicted value is calculated for a Gaussian mixture that models the pdf of y2. Replacing the value of the function g in equations (1) and (2) with the corresponding predicted value of g, the calculated values y2 (k) and y1 (k) of y2 and y1, respectively, become Calculated alternately at each iteration (k):
Initialization:
y1 (0) = x1
Iteration n (n &#8805; 1):
y2 (n) = x2-Ey2 [g (y2, y1, a) | y1 = y1 (n-1)]
y1 (n) = x1-Ey1 [g (y1, y2,1) | y2 = y2 (n)]
n = n + 1

With the principles of the source separation of the present invention generally described above, the source separation process of the present invention in the context of speech recognition will be described.

  Referring first to FIG. 1, a block diagram illustrates the integration of a source separation process in a speech recognition system according to an embodiment of the present invention. As shown, the speech recognition system 100 includes an alignment and scaling module 102, first and second feature extractors 104 and 106, a source separation module 108, a post separation processing module 110, and a speech recognition engine. 112.

  First, in order to compensate for the delay and attenuation introduced during propagation of the signal to a sensor that captures the signal, for example, a microphone (not shown) associated with a speech recognition system, the observed waveform mixtures xpcm1 and xpcm2 are The alignment and scaling module 102 aligns and scales. Such alignment and scaling operations are well known in the art of audio signal processing. Any suitable alignment and scaling techniques can be used.

  Next, cepstral features are extracted from the aligned and scaled waveform mixes xpcm1 and xpcm2 in first and second feature extraction devices 104 and 106, respectively. Techniques for cepstral feature extraction are well known in the field of audio signal processing. Any suitable extraction technique can be used.

  Next, the septal mixtures x1 and x2 output by the feature extractors 104 and 106, respectively, are separated by a source separation module 108 according to the present invention. Obviously, it is desirable that the output of the source separation module 108 is the desired source to which speech recognition is to be applied, for example, in this case the calculated value of the calculated source signal y1. An exemplary source separation process that the source separation module 108 may implement is described in detail below with respect to FIGS.

  The enhanced cepstral features associated with, for example, the calculated source signal y1 output by the source separation module 108 are then normalized and further processed in the post separation processing module 110. Examples of processing techniques that may be performed in the module 110 are also referred to as dynamic features or delta and delta delta cepstral features, where these dynamic features include information about the temporal structure of speech (eg, Rabiner in chapter 3 above). (See references by Mr. et al.), Including calculating its first and second order temporal derivatives and appending it to a vector of cepstral features. , But is not limited thereto.

  Finally, the calculated source signal y1 is sent to the speech recognition engine 112 for decoding. Techniques for performing speech recognition are well known in the field of speech signal processing. Any suitable recognition technique can be used.

  Referring now to FIGS. 2 and 3, there are shown flowcharts of a first portion and a second portion, respectively, of a source isolation process according to an embodiment of the present invention. More specifically, FIGS. 2 and 3, respectively, show two steps forming each iteration of the source separation process according to an embodiment of the present invention.

  First, at time t, the process sets y1 (0, t) equal to the observed mixture x1 (t), i.e., for each time index t, y1 (0, t) = x1 ( Initialized by setting t).

As shown in FIG. 2, the first step 200A of the iteration n (n &#8805; 1) is performed with K Gaussian mixtures N (μ2k, Σ2k) in which the pdf of the random variable y2 has k = 1 to K. By assuming that it is modeled (where N refers to a Gaussian pdf of mean μ2k and difference Σ2k), from the observed mixture x2 and from the calculated value y1 (n-1, t), y1 (0, t) is initialized with x1 (t)) and involves calculating the estimate y2 (n, t) of source y2 at time (t). The steps are represented as follows:
y2 (n, t) = x2 (t) -Σ _k p (k | x2 (t)) g (μ2k, y1 (n-1, t), a) (3)
Note that p (k | x2 (t)) is the random variable x2 following the Gaussian distribution N (μ2k + g (μ2k, y (n-1, t), a), Ξ2k (n, t)) Is calculated in sub-step 202 (post-calculation for Gauss k), where Ξ2k (n, t) is calculated to approximate the difference in random variable x2, where g (u , v, w) = C log (1 + w exp (invC (vu)))). Substep 204 performs a multiplication of p (k | x2 (t)) with g (μ2k, y1 (n-1, t), a), while substep 206 performs x2 (t) and Σ _k p ( k | x2 (t)) g (μ2k, y1 (n-1, t), a) is subtracted. The result is the calculation source y2 (n, t).

As shown in FIG. 3, the second step 200B of the iteration n (n &#8805; 1) is performed with K Gaussian mixtures N (μ1k, Σ1k) in which the pdf of the random variable y1 has k = 1 to K. By assuming to be modeled (where N refers to the Gaussian pdf of the mean μ1k and the difference Σ1k), from the observed mixture x1 and from the calculated value y2 (n, t) at time (t) Includes calculating the calculation y1 (n, t) of source y1. The steps are represented as follows:
y1 (n, t) = x1 (t) -Σ _k p (k | x1 (t)) g (μ1k, y2 (n, t), 1) (4)
Note that p (k | x1 (t)) assumes that the random variable x1 follows the Gaussian distribution N (μ1k + g (μ1k, y2 (n, t), 1), Ξ1k (n, t)) By doing so, it is calculated in sub-step 208 (post-calculation for Gaussian k) (note that Ξ1k (n, t) is calculated to approximate the difference in random variable x1; g (u, v , w) = C log (1 + w exp (invC (vu)))). Substep 210 performs a multiplication of p (k | x1 (t)) and g (μ1k, y2 (n, t), 1), while substep 212 performs x1 (t) and Σ _k p (k | x1 (t)) g (μ1k, y2 (n, t), 1) is subtracted. The result is the calculation source y1 (n, t).

  After M iterations have been performed (M1), a computational stream of T cepstral feature vectors y1 (M, t) for t = 1 to T is sent to the speech recognition engine for decoding. . The computed stream of T cepstral feature vectors y2 (M, t) for t = 1 to T is discarded when it is not decoded. The stream of data y1 is determined to be the source to be decoded based on the relative positions of the microphones capturing streams x1 and x2. A microphone located in close proximity to the audio source to be decoded captures signal x1. A microphone located far from the audio source to be decoded captures the signal x2.

  To elaborate on the above-described exemplary source capture process of the present invention, as pointed out above, the source capture process, at each iteration n, steps 200A and 200B, respectively, of the observed mixtures x1 and x2 used, Calculate the covariance matrix Ξ1k (n, t) or Ξ2k (n, t). The covariance matrix Ξ1k (n, t) or Ξ2k (n, t) is calculated on-the-fly from the observed mixture or is the exponent of the sum of two “log-normally distributed random variables”. It can be calculated according to a Parallel Model Combination (PMC) equation that defines the covariance matrix of the resulting random variables. See, for example, a paper by M.J.F. Gales et al. Entitled "Robust Continuous Speech Recognition Using Parallel Model Combination" in IEEE Transactions on Speechand Audio Processing, vol. 4, 1996.

The PMC equation can be used as follows. Let μ1 and Ξ1 be the average and covariance matrices of the Gaussian random variable z1 in the cepstral domain, respectively. Suppose that μ2 and Ξ2 are the average and covariance matrices of the Gaussian random variable z2 in the cepstral domain, respectively. Assume that z1f = invClog (z1) and z2f = invClog (z2) are random variables obtained by transforming random variables z1 and z2 into the spectral domain. Assume that zf = z1f + z2f is the sum of random variables z1f and z2f. Thus, the PCM equation makes it possible to calculate the covariance matrix の of the random variable z = C log (zf) obtained by transforming the random variable zf into the cepstral domain as follows:
Ξ _ij = log [((Ξ1f _ij + Ξ2f _ij ) / ((μ1f _i + μ2f _i ) (μ1f _j + μ2f _j ))) + 1]
Note that Ξ1f _ij (resp., Ξ2f _ij ) is Ξ1f _ij = μ1f _i * μ1f _j (exp (Ξ1f _ij ) -1) (resp., Ξ2f _ij = μ2f _i * μ2f _j (exp (Ξ2f _ij -1) ) Indicates the (i, j) ^th element in the covariance matrix Ξ1f (resp., Ξ2f), and μ1f _i (resp., Μ2f _i ) indicates the i ^th dimension of the vector μ1f (resp., Μ2f). Μ1f _i = exp (μ1 _i + Ξ1 _ij / 2)) (resp., Μ2f _i = exp (μ2 _i + (Ξ2 _ij / 2))).

  As will be apparent below, in experiments where various spoken voices are mixed with car noise, the pdf of the voice source is modeled with 32 Gaussian mixtures and the pdf of the noise source is , Modeled with two Gaussian mixtures. As far as test data is concerned, 32 Gaussian mixtures for speech and 2 Gaussian mixtures for noise seem to represent a good trade-off between recognition accuracy and complexity. Sources with more complex pdfs may involve more Gaussian mixing.

  Finally, referring to FIG. 4, a block diagram of an exemplary implementation of a speech recognition system incorporating a source separation process (eg, as shown in FIGS. 1, 2 and 3) according to an embodiment of the present invention. Is shown. In this particular implementation 300, a processor 302 for controlling and performing the operations disclosed herein (eg, alignment, scaling, feature extraction, source separation, post-separation processing, and speech recognition) is implemented on a computer. Coupled to memory 304 and user interface 306 via bus 308.

  The term "processor" as used herein is intended to include any processing device, such as, for example, a device including a CPU (central processing unit) and / or other suitable processing circuitry. For example, the processor may be a digital signal processor as known in the art. Also, the term “processor” may refer to a plurality of individual processors. As used herein, the term “memory” refers to a processor or CPU, such as, for example, RAM, ROM, fixed memory devices (eg, hard drives), removable memory devices (eg, floppy disks), etc. It is intended to include memory associated with Further, as used herein, the term "user interface" refers to, for example, a microphone for inputting audio data to a processing unit and, preferably, a visual display device for displaying results associated with a voice recognition process. It is intended to include

  Accordingly, computer software containing instructions or code for performing the methods of the present invention as disclosed herein may comprise one or more associated memory devices (eg, ROM, fixed memory, or removable memory). Memory) and when ready for use, may be partially or fully loaded (eg, into RAM) and executed by the CPU.

  In any event, the elements shown in FIGS. 1, 2 and 3 may be implemented in various forms of hardware, software, or a combination thereof, for example, one or more digital signals having an associated memory. It may be implemented in the form of one or more suitably programmed general purpose digital computers having a processor, application specific integrated circuits, functional circuits, and associated memory. Further, the method of the present invention may be embodied in a machine-readable medium that includes one or more programs that, when executed, implement the steps of the method of the present invention. Given the teachings provided herein regarding the present invention, those skilled in the art will be able to contemplate alternative implementations of the components of the present invention.

  Next, an exemplary evaluation of an embodiment of the present invention used in connection with speech recognition, where the signal mixed with the speech is vehicle noise, will be performed. First, the evaluation protocol is described, after which the recognition scores obtained according to the source separation process of the present invention (hereinafter referred to as “codebook dependent source separation process” or “CDSS”) are obtained. , Is compared to the score obtained without any separation process, and further compared to the score obtained by the MDCCN process described above.

  The experiment is performed on a corpus of twelve male and female subjects emitting a connected digit sequence in a non-moving vehicle. A pre-recorded noise signal in a car at a speed of 60 mph (approximately 96.5 km / hr) is artificially added to the weighted audio signal by a factor of 1 or "a", and therefore has two waveforms: an audio waveform and a noise waveform Different linear mixtures ("ypcm1 + ypcm2" and "aypcm1 + ypcm2" occur as described above, where ypcm1 refers to the speech waveform and ypcm2 refers to the noise waveform). Experiments were performed with coefficient "a" set to 0.3, 0.4, and 0.5. All recordings of speech and noise were made at 22 kHz with an AKG O400 microphone and downsampled to 11 kHz.

  To model the pdf of an audio source, 32 of a collection of thousands of sentences emitted by both men and women and recorded with an AKG Q400 microphone in a non-moving car and noise-free environment were used. Gaussian mixture was calculated. Approximately 4 minutes of noise recorded with an AKG Q400 in a 60 mph (approximately 96.5 km / hr) vehicle using the same settings as for the test data to model the pdf of the vehicle noise For each (prior to the experiment) two Gaussian mixtures were calculated.

The mixture of speech and noise decoded by the speech recognition engine is
(A) not separated, or (B) separated by MDCCN process, or (C) separated by CDSS process.
The performance of the speech recognition engine obtained by the above (A), (B) and (C) is compared by Word Error Rates (WER).

  The speech recognition engine used in that experiment is used in particular in portable devices or in automotive applications. The engine was trained with about 10,000 context-dependent Gaussian, or general English, voices for hundreds of hours (approximately half of these training data digitally added car noise, Or a triphone context united by using a decision tree (recorded in a car traveling at speeds of 30 mph and 60 mph (about 48 km / hour and about 96.5 km / hour)). And a set of speaker independent acoustic models (156 subphones covering English voice). (See the paper by LR Bahl et al., Entitled "Performance of the IBM Large Vocabulary Continuous Speech Recognition System on the ARPA Wall Street JournalTask," in Proceedings of ICASSP 1995, vol. 1, pp. 41-44. want). The front end of the system computes twelve Sepstra + energy + delta and delta-delta coefficients from a 15 ms frame using a 24 mel filter bank (see, for example, Rabiner et al., Chapter 3, supra). Please see).

  The CDSS process is generally applied as described above, and is preferably applied as exemplified above in connection with FIGS. 1, 2, and 3.

  Table 1 below shows the word error rate (WER) obtained after decoding the test data. The WER obtained for a clean speech before adding noise is 1.53%. The WER obtained in noisy speech after the addition of noise and without using any separation process is 12.31%. The WER obtained after using the MCDCN process using the second mixture ("ayf1 + yf2") as the reference signal is given for various values of the mixing coefficient "a". The MCDCN provides a reduction in WER when the speech leakage in the reference signal is small (a = 0.3), but its performance decreases as the leakage becomes more important, with a coefficient "a" equal to 0.5. , The MDCCN process is worse than the 12.31% baseline WER. On the other hand, the CDSS process greatly improves the baseline WER for all experimental values of the coefficient "a".

(Table 1)
Original sound 1.53
Noisy voice, no separation 12.31
a = 0.3 a = 0.4 a = 0.5
Noisy voice, MCDCN 7.86 10.00 15.51
Noisy voice, CDSS 6.35 6.87 7.59

  Although the embodiments of the present invention have been described with reference to the accompanying drawings, the present invention is not limited to the embodiments themselves, and various other changes and modifications can be made without departing from the scope or spirit of the present invention. Can be performed by those skilled in the art.

  In summary, the following matters are disclosed regarding the configuration of the present invention.

(1) A method for separating a signal from a mixture of a signal related to a first source (a first source signal) and a signal related to a second source (a second source signal),
Obtaining two signals, each representing two mixtures of the first source signal and the second source signal;
Using the two signals and at least one known statistical property associated with the first source and the second source and without using a reference signal from the mixture in the non-linear signal domain Separating the first source signal;
A method that includes
(2) The above (1), wherein the two signals respectively represent an unweighted mixed signal of the first source signal and the second source signal and a weighted mixed signal of the first source signal and the second source signal. ).
(3) the step of separating is performed in the non-linear domain by converting the unweighted mixed signal into a first cepstral mixed signal and converting the weighted mixed signal into a second cepstral mixed signal. The method according to (2).
(4) the separating step iteratively calculates a calculated value for the second source signal based on the second cepstral mixed signal and a calculated value for the first source signal from a previous iteration in the separating step. The method according to (3), further comprising the step of generating.
(5) The method according to (4), wherein the step of generating an estimate for the second source signal assumes that the second source signal is modeled by Gaussian mixing.
(6) The step of separating further includes a step of repeatedly generating a calculated value for the first source signal based on the calculated value for the first cepstral mixed signal and the calculated value for the second source signal. The method according to 4).
(7) The method according to (6), wherein the step of generating an estimate for the first source signal assumes that the first source signal is modeled by Gaussian mixing.
(8) The method according to (1), wherein the separated first source signal is subsequently used by a signal processing application.
(9) The method according to (8), wherein the application is speech recognition.
(10) The above (1), wherein the first source signal is an audio signal, and the second source signal is a signal representing at least one of competing audio, interfering music, and a specific noise source. the method of.
(11) An apparatus for separating a signal from a mixture of a signal related to a first source (first source signal) and a signal related to a second source (second source signal),
Memory and
Coupled to the memory, operative to obtain two mixed signals respectively representing two bodies of the first source signal and the second source signal; and (ii) the two signals and the first source signal. And separating the first source signal from the mixture in the non-linear signal domain using at least one known statistical property associated with the second source and without requiring the use of a reference signal. At least one processor operating on
Equipment including.
(12) The above (11), wherein the two signals respectively represent an unweighted mixed signal of the first source signal and the second source signal and a weighted mixed signal of the first source signal and the second source signal. The device according to (1).
(13) the separating operation is performed in the non-linear domain by converting the non-weighted mixed signal into a first cepstral mixed signal and converting the weighted mixed signal into a second cepstral mixed signal; The device according to the above (12).
(14) the separating operation iteratively generating a calculated value for the second source signal based on the second cepstral mixed signal and a calculated value for the first source signal from a previous iteration of the separating operation. The apparatus according to the above (13), including an operation of performing the following.
(15) The apparatus according to (14), wherein the operation of generating an estimate for the second source signal assumes that the second source signal is modeled by Gaussian mixing.
(16) The method according to (14), wherein the separating operation further includes an operation of repeatedly generating a calculated value for the first source signal based on the calculated value for the first cepstral mixed signal and the second source signal. The device according to (1).
(17) The apparatus according to (16), wherein the operation of generating an estimate for the first source signal assumes that the first source signal is modeled by Gaussian mixing.
(18) The apparatus according to (11), wherein the separated first source signal is subsequently used by a signal processing application.
(19) The device according to (18), wherein the application is voice recognition.
(20) The above (11), wherein the first source signal is an audio signal, and the second source signal is a signal representing at least one of competing audio, interfering music, and a specific noise source. Equipment.
(21) A computer program for separating a signal from a mixture of a signal related to a first source (a first source signal) and a signal related to a second source (a second source signal),
Obtaining two signals, each representing two mixtures of the first source signal and the second source signal;
Using the two signals and at least one known statistical property associated with the first source and the second source and without using a reference signal from the mixture in the non-linear signal domain Separating the first source signal;
A computer program comprising a machine-readable medium including one or more programs that implement at runtime.
(22) The above (21), wherein the two signals respectively represent an unweighted mixed signal of the first source signal and the second source signal and a weighted mixed signal of the first source signal and the second source signal. The computer program according to (1).
(23) the separating is performed in the non-linear domain by converting the unweighted mixed signal into a first cepstral mixed signal and converting the weighted mixed signal into a second cepstral mixed signal; The computer program according to the above (22).
(24) the separating step iteratively generating a calculated value for the second source signal based on the second cepstral mixed signal and a calculated value for the first source signal from a previous iteration of the separating step. The computer program according to the above (23), comprising the step of:
(25) The computer program according to (24), wherein the step of generating an estimate for the second source signal assumes that the second source signal is modeled by Gaussian mixture.
(26) The above (24), wherein the step of separating further includes the step of repeatedly generating a calculated value for the first source signal based on the calculated value for the first cepstral mixed signal and the second source signal. The computer program according to (1).
(27) The computer program according to (26), wherein the step of generating an estimate for the first source signal assumes that the first source signal is modeled by Gaussian mixture.
(28) The computer program according to (21), wherein the separated first source signal is subsequently used by a signal processing application.
(29) The computer program according to (28), wherein the application is voice recognition.
(30) The above (21), wherein the first source signal is an audio signal, and the second source signal is a signal representing at least one of competing audio, interfering music, and a specific noise source. Computer programs.
(31) An apparatus for separating a signal from a mixture of a signal related to a first source (first source signal) and a signal related to a second source (second source signal),
Means for obtaining two signals respectively representing two mixtures of said first source signal and said second source signal;
Coupled to the means for obtaining the two signals, using the two signals and at least one known statistical property associated with the first source and the second source and requiring the use of a reference signal Means for separating the first source signal from the mixture in the non-linear signal domain, and
Equipment including.

FIG. 4 is a block diagram illustrating integration of a source separation process in a speech recognition system according to an embodiment of the present invention. 5 is a flowchart illustrating a first part of a source separation process according to an embodiment of the present invention. 5 is a flowchart illustrating a second part of the source separation process according to an embodiment of the present invention. FIG. 2 is a block diagram illustrating an exemplary implementation of a speech recognition system incorporating a source separation process according to an embodiment of the present invention.

Claims (31)

A method for separating a signal from a mixture of a signal associated with a first source (a first source signal) and a signal associated with a second source (a second source signal),
Obtaining two signals, each representing two mixtures of the first source signal and the second source signal;
Using the two signals and at least one known statistical property associated with the first source and the second source and without using a reference signal from the mixture in the non-linear signal domain Separating the first source signal;
A method that includes   2. The signal of claim 1, wherein the two signals represent an unweighted mixed signal of the first source signal and the second source signal and a weighted mixed signal of the first source signal and the second source signal, respectively. 3. Method.   3. The method of claim 2, wherein the separating is performed in the non-linear domain by converting the non-weighted mixed signal to a first cepstral mixed signal and converting the weighted mixed signal to a second cepstral mixed signal. The described method.   The separating step iteratively generating a calculated value for the second source signal based on the second cepstral mixed signal and a calculated value for the first source signal from a previous iteration of the separating step. 4. The method of claim 3, comprising:   The method of claim 4, wherein generating an estimate for the second source signal assumes that the second source signal is modeled by Gaussian mixing.   5. The method of claim 4, wherein the separating step further comprises: iteratively generating a calculated value for the first source signal based on the first cepstral mixed signal and a calculated value for the second source signal. the method of.   7. The method of claim 6, wherein generating an estimate for the first source signal assumes that the first source signal is modeled by Gaussian mixing.   The method of claim 1, wherein the separated first source signal is subsequently used by a signal processing application.   The method of claim 8, wherein the application is speech recognition.   The method of claim 1, wherein the first source signal is an audio signal and the second source signal is a signal representing at least one of competing audio, interfering music, and a particular noise source. An apparatus for separating a signal from a mixture of a signal associated with a first source (a first source signal) and a signal associated with a second source (a second source signal),
Memory and
Coupled to the memory, operable to obtain two signals, each representing two mixtures of the first source signal and the second source signal, and (ii) the two signals and the first source signal. And separating the first source signal from the mixture in the non-linear signal domain using at least one known statistical property associated with the second source and without requiring the use of a reference signal. At least one processor operating on
Equipment including.   The method of claim 11, wherein the two signals represent a non-weighted mixed signal of the first and second source signals and a weighted mixed signal of the first and second source signals, respectively. apparatus.   13. The method of claim 12, wherein the separating is performed in the non-linear domain by converting the non-weighted mixed signal to a first cepstral mixed signal and converting the weighted mixed signal to a second cepstral mixed signal. An apparatus according to claim 1.   The separating operation includes an iteratively generating a calculated value for the second source signal based on the second cepstral mixed signal and a calculated value for the first source signal from a previous iteration of the separating operation. 14. The device of claim 13, comprising.   15. The apparatus of claim 14, wherein generating an estimate for the second source signal assumes that the second source signal is modeled by Gaussian mixing.   The method of claim 14, wherein the separating operation further comprises: iteratively generating an estimate for the first source signal based on the estimate for the first cepstral mixed signal and the second source signal. apparatus.   17. The apparatus of claim 16, wherein generating an estimate for the first source signal assumes that the first source signal is modeled by Gaussian mixing.   The apparatus of claim 11, wherein the separated first source signal is subsequently used by a signal processing application.   19. The device of claim 18, wherein the application is speech recognition.   The apparatus of claim 11, wherein the first source signal is an audio signal and the second source signal is a signal representing at least one of competing audio, interfering music, and a particular noise source. A computer program for separating a signal from a mixture of a signal associated with a first source (a first source signal) and a signal associated with a second source (a second source signal),
Obtaining two signals, each representing two mixtures of the first source signal and the second source signal;
Using the two signals and at least one known statistical property associated with the first source and the second source and without using a reference signal from the mixture in the non-linear signal domain Separating the first source signal;
A computer program comprising a machine-readable medium including one or more programs that implement at runtime.   22. The signal of claim 21, wherein the two signals respectively represent an unweighted mixed signal of the first source signal and the second source signal and a weighted mixed signal of the first source signal and the second source signal. Computer program.   23. The separating step is performed in the non-linear domain by converting the unweighted mixed signal to a first cepstral mixed signal and converting the weighted mixed signal to a second cepstral mixed signal. A computer program according to claim 1.   The separating step comprises the step of iteratively generating a calculated value for the second source signal based on the second cepstral mixed signal and a calculated value for the first source signal from a previous iteration of the separating step. 24. The computer program according to claim 23, comprising:   26. The computer program of claim 24, wherein generating an estimate for the second source signal assumes that the second source signal is modeled by Gaussian mixing.   25. The method of claim 24, wherein the separating further comprises: iteratively generating a calculated value for the first source signal based on the calculated value for the first cepstral mixed signal and the second source signal. Computer program.   27. The computer program of claim 26, wherein generating an estimate for the first source signal assumes that the first source signal is modeled by Gaussian mixing.   22. The computer program of claim 21, wherein the separated first source signal is subsequently used by a signal processing application.   29. The computer program according to claim 28, wherein the application is speech recognition.   22. The computer program of claim 21, wherein the first source signal is a speech signal and the second source signal is a signal representing at least one of competing speech, interfering music, and a particular noise source. . An apparatus for separating a signal from a mixture of a signal associated with a first source (a first source signal) and a signal associated with a second source (a second source signal),
Means for obtaining two signals respectively representing two mixtures of said first source signal and said second source signal;
Coupled to the means for obtaining the two signals, using the two signals and at least one known statistical property associated with the first source and the second source and requiring the use of a reference signal Means for separating the first source signal from the mixture in the non-linear signal domain, and
Equipment including.

Images